Effect of voice disguise on the performance of a forensic automatic speaker recognition system

This paper presents first results of an ongoing study on the effects of common types of voice disguise, including increased voice pitch (even falsetto speech), lowered voice pitch and pinching the nose while speaking, on forensic speaker recognition (FSR) techniques. Natural and disguised speech data from 100 German speakers recorded 5 times over a period of 7 to 9 months were used in a series of speaker recognition experiments, using the LR-based forensic automatic speaker recognition system developed by ATVS at Universidad Politécnica de Madrid. In this paper, experiments are limited to estimate the performance degradation when the suspect is known to be the author of the disguised test speech (no impostor trials are reported). Results indicate that the three types of voice disguise selected affect the performance of the system only marginally if reference populations contain speech data which exhibit the same type of disguise. If, however, the reference population is assembled with normal speech only, effects are generally more severe and also different for the three types of disguise under evaluation.